Open Weather Data Challenge Runner-Up Predicts Flash Floods
The Vaisala Open Weather Data Challenge competition challenged students and professionals to come up with creative and innovative ways to utilize open weather data. The Global Flash Flood Prediction site, by Zac Flamig and Race Clark, was one of the runner-ups in the challenge. Here they discuss their concept and its impact.
Zac Flamig’s research interests include hydrologic modeling at small scales over large domains, crowd sourced data collection, disaster response and radar meteorology. When Zac isn’t working on hydrologic models he enjoys developing apps for iOS, chasing storms in Oklahoma and following the latest space launch news. Race (Robert) Clark’s primary research focuses on applying data mining and machine learning to NWP outputs to forecast hydrometeorological hazards, particularly flash flooding. When Race is not at work or doing campus activities, he enjoys hiking, kayaking, chasing Oklahoma’s storms, and cheering on the Oklahoma City Thunder.
We had been working in hydrologic modeling for several years and had explored developing tools for forecasting and monitoring floods, so the Open Weather Data Challenge was a good forum to get these ideas out in front of a new audience. Our faculty advisor at the University of Oklahoma encouraged us to enter. Before the Challenge, we had had smaller projects using hydrologic models in e.g. southwestern and east Africa, and the US. The Challenge gave us the opportunity to unify all these into one global project.
The Global Flash Flood Prediction page is actually a combination of two sub-projects. Zac has been lead developer in a project to model flash flooding over the United States, using the Ensemble Framework for Flash Flood Forecasting (http://ef5.ou.edu ). This methodology was extended globally for this project by changing how we used a few key datasets. For instance, in the US, rainfall provided as input to the hydrologic model comes from a network of ground weather radars. Globally, ground radar is not as widely available, so we turned to satellite-based estimates of rainfall.
This project also gave us the opportunity to use 72-hour global forecast rainfall as an input to our model. This extends the effective lead-time of the flooding predictions and hopefully makes them more useful to decision makers. We wanted also to provide somewhat independent predictions of flooding environments from the atmospheric perspective. So we also used a data mining and machine learning technique – the random forest – applied to archived GFS weather model output to forecast the probability of flash flooding events around the world. This is closely related to Race’s Ph.D. dissertation research.
It was important to us as well to provide some verification of the predictions. In the US, we used automated reports from the U.S. Geological Survey’s network of river gauges. Globally, we display flooding reports from the Dartmouth Flood Observatory’s RiverWatch system. Of course, it was important that all the data used in the project be open, but in keeping with the spirit of the Challenge we also chose open-source software and free processing tools.
The site uses open data from all sorts of sensors and projects: numerical weather prediction outputs, flood reports (both human-augmented and automatically-collected), satellite rainfall estimates, global topography, global soils, climatological potential evapotranspiration, and more. We foresee this project being of use in developing regions, so the site is quick to load and requires relatively small bandwidth. We update our hydrologic model outputs every 3 hours and our random forest predictions every six hours.
We would anticipate forecasters, hydrologists, emergency response officials, and others being able to use our site to assess the 3-day flooding threat for their area and plan a response accordingly. In many areas of the world right now, flooding is not something to anticipate, but something that you only respond to once it has already occurred. This is not due to a lack of scientific understanding, but to a lack of easily-available and interpretable data. By combining the necessary datasets into one single easy-to-use site, we can help improve outcomes and reduce the impacts from floods around the world.
Zac and Race are Ph.D. candidates in the School of Meteorology at the University of Oklahoma and work as Research Associates for OU's Cooperative Institute for Mesoscale Meteorological Studies at the National Severe Storms Laboratory.